Final Project
Data Visualization (STAT 302)
1 Introduction
1.1 Objective
In their 2018 article “Investor Rights versus Human Rights: Do Bilateral Investment Treaties Tilt the Scale?”, Bodea & Ye find that bilaterial investment treaties (BITs) have, on balance, a negative effect on human rights respect in developing countries. The pith of their argument is that—particularly in states with weak democratic institutions—BITs undermine working conditions and social welfare, inviting popular anti-government activities; these instances of dissent are in turn met with governmental repression. Motivating the second chapter of my dissertation is my desire to test this claim.
In writing this chapter, I hope to use spatial-modelling methods1 on my main dependent and independent variables. However, before I commit to doing so, I would like to justify the need for such an approach by assessing the extent of spatial clustering therein. I seek here to accomplish this very task, mainly by means of data visualization.
1 Examples of such potential methods include spatial autoregressive models (SARs), spatial error models (SEMs), spatially-lagged X models (SLX), spatial Durbin models (SDMs), spatial Durbin error models (SDEMs), combined spatial autocorrelation models (SACs), and general nesting spatial models (GNSs). For a helpful summary of said models, see “Spatial Data Analysis” (Rüttenauer, 2024).
1.2 Memo Outline
My memo proceeds as follows. To start (Section 2), synthesizing rnaturalearth spatial data and the main variables of interest, as well as using gganimate, I provide world-map animations depicting time trends for these variables, namely:
- Human Rights (HR) Scores (Christopher Fariss, University of Michigan), a single variable measuring the quality of physical integrity rights respect by country-year.
- Three measures of cumulative BITs taken from Bodea & Ye’s replication data.
- A measure of “political dissent” found in the foregoing replication data but recreated from the Cross-National Time-Series Dataset (CNTS) to include more recent observations.
These animations allow me to comment on whether reason exists to suspect the presence of spatial clustering. Then (Section 3), using sf and spdep, I compute and visualize Global Moran’s \(I\) for each main variable to furnish yet more evidence in favor of or against the existence of spatial clustering. I conclude (Section 4) with a commentary on my findings writ large, suggesting whether and where spatial modelling is appropriate for my research.
2 Animations
2.1 Human Rights Scores
Unlike Bodea & Ye, who rely on the CIRI index of physical integrity rights, I intend for Fariss’s HR Scores to serve as my main dependent variable. Figure 1, below, gives HR scores by country over the duration of the dataset (1946-2019):
Countries/geographical objects with missing data are shaded in grey, but those possessing HR scores are shaded along a red-yellow-green continuum, with deep red representing the “worst” human-rights offenders, and deep green representing “best” human-rights respecters.2 With this in mind, that the map generally becomes less “red” (i.e., “yellower” and “greener”) as it approaches the present lends obvious support to Fariss’s (2014) central claim: human-rights respect is improving over time.
2 That is, red signifies a negative HR score, whereas green symbolizes a positive one.
More pertinent to our interests is the observation that human-rights respect does appear regionally clustered over the years. In particular, human rights “respecters” seem congregated in Western Europe, North America, Oceania, East Asia, and western South America; meanwhile, human rights “offenders” seem more commonplace in Africa (sans Namibia and Botswana) as well as West, Central, and South Asia. For obvious reasons, this distribution largely mirrors that of democracies versus autocracies—a geographical spread of which most are aware—so this “finding” is unlikely to come as a surprise. Still, Figure 1 is useful insofar as it suggests that the data support our intuitions regarding the inherent spatial clustering of human rights respect.
2.2 Bilateral Investment Treaties
2.2.1 Simple Count
The most basic treatment variable that Bodea & Ye deploy is a count of the cumulative number of BITs ratified per country-year:
Vis-à-vis HR Scores, the scope of the BITs data is more restrictive in that (1.) the country cases are limited to “developing” ones, of which the authors identify 113;3 and (2.) it covers the narrower time span of 1958-2009. As a result, the animation is shorter and contains more “missing” (grey) values. However, like Figure 1, Figure 2 appears to evince spatial clustering: in particular, and especially beginning ca. 1990, we see the cumulative number of BITs accelerating (represented by darkening shades of orange) in Eastern Europe, East/South Asia, North Africa, and western South America. This may well lend credence to another of Bodea & Ye’s hypotheses: that a country’s ratification of BITs may exert competitive pressures on neighboring countries to ratify BITs themselves.
3 According to Bodea & Ye, “these are countries that the World Bank does not classify as high income for the majority of our sample period” (2018, p. 964).
2.2.2 ICSID
A variant of the standard cumulative BITs count is one that exclusively tallies BITs containing an enforcement mechanism through the International Centre for Settlement of Investment Disputes (ICSID):
In Figure 3, above, we see some degree of clustering, namely in Southeast Asia, the Balkans, the Pacific coast of South America, and (to a lesser extent) the “neighborhood” centered on the Arabian Peninsula. However, this clustering appears less prominent, especially relative to that seen in Figure 2. This may result from countries being more hesitant to accede to enforceable treaties as a general matter.
2.2.3 North-South BITs
The final main variant of the cumulative BITs count is that which sums the number of BITs entailing a “Global North-Global South” relationship:
The trends shown in Figure 4 are highly similar to those seen in Figure 2, including the appearance of spatial clustering in the regions mentioned in Section 2.2.1. Perhaps this is the case in virtue of BITs generally entailing a North-South relationship.
2.3 Political Dissent
Following Nordas & Davenport (2013), Bodea & Ye code political dissent “as the sum of antigovernment protest, riots and general strikes”; these constitutive variables are found in the CNTS dataset, as aforementioned in Section 1.2 (2018, p. 966). Recreating the variable to capture years up to 2023 and for all countries yields the following animation:
As we can see, Figure 5 is uninformative with respect to detecting spatial clustering; the outlier of the United States in 2020 is so extreme that the color scale is effectively broken. However, it does draw our attention to the problem of instrumentation bias, since it is extremely unlikely that the amount of protests witnessed in the United States that year, though undoubtedly high, far exceeded those for any other country-year by many orders of magnitude. And indeed, the CNTS user manual4 acknowledges that data taken prior to 2011 relied mainly on articles appearing in The New York Times, whereas more recent data have come from a wider pool of sources (e.g., credible news stories on the Internet). In toto, then, the political dissent variable is biased towards events deemed relevant to English-language (and largely American) audiences, and towards more recent anti-government activities, whose very existence may register more readily outside of “traditional” channels.
4 See data/cnts_data_2024/2024_edition_user_manual.docx
However, it is fair to wonder whether said instrumentation bias is at work, or is just as acute, for the observations in Bodea & Ye’s dataset, which (as aforementioned) doesn’t feature developed countries such as the United States. We can therefore recreate Figure 5 by limiting the observations to exclusively those in the foregoing dataset, yielding us the following:
The scale seen in Figure 6, above, is about 1/10th the width of that seen in Figure 5, meaning we can actually perceive variation in instances of popular dissent over time aside from one country-year. Nevertheless, a bias towards more recent (and perhaps English language-relevant) observations seems to remain, with more instances of said dissent being counted in recent years and in India, specifically. That being said, Figure 6 is useful inasmuch as we seem to see a lack of spatial clustering: contrary to expectations, perhaps, numbers of antigovernment incidents appear to wax and wane by country, not by region.
3 Global Moran’s \(I\)
The animations above have furnished us with preliminary evidence for the presence—or lack—of spatial clustering in our main variables of interest. However, these suspicions may be rigorously evaluated by computing and visualizing the “Moran’s \(I\)” statistic over time. In brief, Moran’s \(I\) seeks to convey the extent of spatial autocorrelation in a given dataset. The statistic spans the interval [-1, 1]. With respect to the variable of interest, values closer to 1 signify near-perfect clustering, while values approaching -1 entail near-perfect dispersal/opposition. Meanwhile, values near 0 suggest a completely random spatial distribution.5 Importantly, computing Moran’s \(I\) requires a weighted neighborhood matrix, wherein a criterion for neighbor membership (e.g., a certain distance between two points) must be specified.
5 For more on Moran’s \(I\), including its formula, see its Wikipedia article.
6 This generally follows Bodea & Ye (2018, p. 968)
7 This is a more capacious criterion supplied by Ward & Gleditsch (2007, p. 12).
For each of the variables analyzed in Section 2, I have computed their “global” (i.e., full dataset-encompassing) Moran’s \(I\) per year and based on two neighborhood criteria: whether the country polygons comprising the shape data are within 12 miles at their nearest points,6 or whether these points are within 200 kilometers of each other.7 These statistics, as well as their accompanying p-values, are visualized in Figure 7, Figure 8, Figure 9, Figure 10, and Figure 11, below:
In each case, changing the specification for neighborhood membership does little to alter Moran’s \(I\), meaning our findings seem rather robust. More importantly, however, is that in every instance, the Moran’s \(I\) statistics and their accompanying p-values accord with the findings we made more informally in Section 2: spatial clustering is apparent in HR scores (Figure 1 & Figure 7), as well as in cumulative ratified BITs (Figure 2 & Figure 8) and North-South BITs (Figure 4 & Figure 10).8 Meanwhile, the extent of spatial clustering between cumulative ratified BITs and North-South BITs appears generally the same.9 However, we see less clustering in ICSID-enforced BITs (Figure 3 & Figure 9),10 and the distribution of instances of political dissent is clearly distributed randomly (Figure 6 & Figure 11).11
8 In each case, Moran’s \(I\) approaches 0.4-0.5, and the p-values bottom out near 0.
9 Indeed, Moran’s \(I\) and the p-values thereof are virtually the same across the years.
10 In particular, the Moran’s \(I\) statistic, though statistically significant, plateaus at a more modest 0.2.
11 Indeed, Moran’s \(I\) and its accompanying p-value fluctuate wildly over time.
4 Conclusion
Taken together, our findings in Section 2 & Section 3 suggest that the variables best suited to spatial clustering methods are HR Scores and cumulative ratified BITs/North-South BITs. Having made such findings, I believe that this exercise has been invaluable in helping me ascertain what variables—if any—may be deployed for spatial modelling for the second chapter of my dissertation. Moving forward, I hope to interrogate the “political dissent” variable, specifically, and determine whether the data may be modified to better capture the presence and/or spread of anti-government activities, for example by generating a binary indicator for the presence or lack of such activities per year.